Non Homogeneous Realistic Single Image Dehazing

Abstract

A hazy image is one where atmospheric effects degrade the contrast and visibility of the image. It is often caused by the dispersion of light into the moisture particles present, smoke etc. This results in lower performance in high level vision tasks such as object detection, free space detection, scene understanding, etc. Hence the images have to be de-hazed before applying other high level algorithms. Dehazing is the process of reconstructing the original colour and contrast of the image if taken in normal conditions. Image dehazing is a non-trivial task as it is hard to collect haze free ground truth images. Further, achieving dehazed images when variable haze is present is a significantly harder challenge. In this research, we propose the Non Homogeneous RESIDE dataset (NH-RESIDE) that contains images created synthetically using the principles of randomness and representativeness. Experimental results show that the model trained on our dataset produces visually more pleasing images with a much better dehazing effect on real world images. The model implemented in this paper also outperforms the state-of-the-art models by a huge margin on the NH-Haze dataset proposed by the NTIRE Non Homogeneous Dehazing Challenge at CVPR, achieving an average PSNR of 25.69 and an average SSIM of 0.80. It also achieves much better processing times when compared to other models, thereby facilitating real-time performance.